|Publication number||US7996219 B2|
|Application number||US 11/085,587|
|Publication date||Aug 9, 2011|
|Filing date||Mar 21, 2005|
|Priority date||Mar 21, 2005|
|Also published as||CA2540557A1, US8731924, US20060212293, US20110295602, US20140330565|
|Publication number||085587, 11085587, US 7996219 B2, US 7996219B2, US-B2-7996219, US7996219 B2, US7996219B2|
|Original Assignee||At&T Intellectual Property Ii, L.P.|
|Export Citation||BiBTeX, EndNote, RefMan|
|Patent Citations (17), Non-Patent Citations (1), Referenced by (1), Classifications (4), Legal Events (3)|
|External Links: USPTO, USPTO Assignment, Espacenet|
1. Field of the Invention
The present invention relates to speech processing and more specifically to adapting an existing language model to a new natural language spoken dialog application.
Natural language spoken dialog systems receive spoken language as input, analyze the received spoken language input to derive meaning from the input, and perform some action, which may include generating speech, based on the meaning derived from the input. Building natural language spoken dialog systems requires large amounts of human intervention. For example, a number of recorded speech utterances may require manual transcription and labeling for the system to reach a useful level of performance for operational service. In addition, the design of such complex systems typically includes a human being, such as a User Experience (UE) expert to manually analyze and define system core functionalities, such as, a system's semantic scope (call-types and named entities) and a dialog manager strategy, which will drive the human-machine interaction. This approach to building natural language spoken dialog systems is extensive and error prone because it involves the UE expert making non-trivial design decisions, the results of which can only be evaluated after the actual system deployment. Thus, a complex system may require the UE expert to define the system's core functionalities via several design cycles which may include defining or redefining the core functionalities, deploying the system, and analyzing the performance of the system. Moreover, scalability is compromised by time, costs and the high level of UE know-how needed to reach a consistent design.
Additional features and advantages of the invention will be set forth in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. The features and advantages of the invention may be realized and obtained by means of the instruments and combinations particularly pointed out in the appended claims. These and other features of the present invention will become more fully apparent from the following description and appended claims, or may be learned by the practice of the invention as set forth herein.
In a first aspect of the invention, a method is provided for building a spoken language understanding model. Labeled data are obtained for a target application. A new classification model is formed for use with the target application by using the labeled data for adaptation of an existing classification model.
In a second aspect of the invention, an apparatus is provided. The apparatus includes a processor and storage for storing instructions for the processor. The apparatus is configured to obtain labeled data for a target application, and form a new classification model for use with the target application by using the labeled data for adaptation of an existing classification model.
In a third aspect of the invention, a machine-readable medium having instructions, stored therein, for a processor is provided. The machine-readable medium includes instructions for inputting labeled data for a target application, and instructions for forming a new classification model for use with the target application by using the labeled data for adaptation of an existing classification model.
In a fourth aspect of the invention, an apparatus is provided. The apparatus includes means for obtaining labeled data for a target application, and means for forming a new classification model for use with the target application by using the labeled data for adaptation of an existing classification model.
In order to describe the manner in which the above-recited and other advantages and features of the invention can be obtained, a more particular description of the invention briefly described above will be rendered by reference to specific embodiments thereof which are illustrated in the appended drawings. Understanding that these drawings depict only typical embodiments of the invention and are not therefore to be considered to be limiting of its scope, the invention will be described and explained with additional specificity and detail through the use of the accompanying drawings in which:
Various embodiments of the invention are discussed in detail below. While specific implementations are discussed, it should be understood that this is done for illustration purposes only. A person skilled in the relevant art will recognize that other components and configurations may be used without parting from the spirit and scope of the invention.
Spoken dialog systems aim to identify intents of humans, expressed in natural language, and take actions accordingly, to satisfy their requests.
ASR module 102 may analyze speech input and may provide a transcription of the speech input as output. SLU module 104 may receive the transcribed input and may use a natural language understanding model to analyze the group of words that are included in the transcribed input to derive a meaning from the input. The role of DM module 106 is to interact in a natural way and help the user to achieve the task that the system is designed to support. DM module 106 may receive the meaning of the speech input from SLU module 104 and may determine an action, such as, for example, providing a response, based on the input. SLG module 108 may generate a transcription of one or more words in response to the action provided by DM 106. TTS module 110 may receive the transcription as input and may provide generated audible speech as output based on the transcribed speech.
Thus, the modules of system 100 may recognize speech input, such as speech utterances, may transcribe the speech input, may identify (or understand) the meaning of the transcribed speech, may determine an appropriate response to the speech input, may generate text of the appropriate response and from that text, may generate audible “speech” from system 100, which the user then hears. In this manner, the user can carry on a natural language dialog with system 100. Those of ordinary skill in the art will understand the programming languages and means for generating and training ASR module 102 or any of the other modules in the spoken dialog system. Further, the modules of system 100 may operate independent of a full dialog system. For example, a computing device such as a smartphone (or any processing device having a phone capability) may have an ASR module wherein a user may say “call mom” and the smartphone may act on the instruction without a “spoken dialog.”
Processor 220 may include at least one conventional processor or microprocessor that interprets and executes instructions. Memory 230 may be a random access memory (RAM) or another type of dynamic storage device that stores information and instructions for execution by processor 220. Memory 230 may also store temporary variables or other intermediate information used during execution of instructions by processor 220. ROM 240 may include a conventional ROM device or another type of static storage device that stores static information and instructions for processor 220. Storage device 250 may include any type of media, such as, for example, magnetic or optical recording media and its corresponding drive.
Input device 260 may include one or more conventional mechanisms that permit a user to input information to system 200, such as a keyboard, a mouse, a pen, a voice recognition device, etc. Output device 270 may include one or more conventional mechanisms that output information to the user, including a display, a printer, one or more speakers, or a medium, such as a memory, or a magnetic or optical disk and a corresponding disk drive. Communication interface 280 may include any transceiver-like mechanism that enables system 200 to communicate via a network. For example, communication interface 280 may include a modem, or an Ethernet interface for communicating via a local area network (LAN). Alternatively, communication interface 280 may include other mechanisms for communicating with other devices and/or systems via wired, wireless or optical connections. In some implementations of natural spoken dialog system 100, communication interface 280 may not be included in processing system 200 when natural spoken dialog system 100 is implemented completely within a single processing system 200.
System 200 may perform such functions in response to processor 220 executing sequences of instructions contained in a computer-readable medium, such as, for example, memory 230, a magnetic disk, or an optical disk. Such instructions may be read into memory 230 from another computer-readable medium, such as storage device 250, or from a separate device via communication interface 280.
Boosting is an iterative procedure; on each iteration, t, a weak classifier, h, is trained on a weighted training set, and at the end of training, the weak classifiers are combined into a single, combined classifier. The algorithm generalized for multi-class and multi-label classification is shown in
The algorithm may begin by initializing a uniform distribution D1 (i, l) over training examples i and labels l. After each round this distribution may be updated so that the example-class combinations, which are easier to classify, get lower weights and vice versa. The intended effect is to force the weak learning algorithm to concentrate on the examples and labels that will be the most beneficial to the overall goal of finding a highly accurate classification rule.
This algorithm can be seen as a procedure for finding a linear combination of base classifiers which attempts to minimize an exponential loss function, which in this case is:
An alternative would be to minimize a logistic loss function, namely
In that case, the confidence of a class, l, for an example, xi may be computed as:
A more detailed explanation and analysis of this algorithm can be found in R. E. Schapire, “The boosting approach to machine learning: An overview,” in Proceedings of the ICAASP, Hong Kong, April 2003, which is incorporated by reference herein in its entirety. In experiments, a BoosTexter tool, which is an implementation of the Boosting algorithm, was used. For text categorization, BoosTexter uses word n-grams as features, and each weak classifier (or “decision stump”) checks the absence or presence of a feature.
Implementations consistent with the principles of the invention may exploit existing labeled data and models for boosting the performance of new similar applications using a supervised adaptation method. The basic assumption is that there is an intent model trained with data similar to the target application. This classification model may be adapted using a small amount of already labeled data from the target application, thus reducing the amount of human-labeling effort necessary to train decent statistical intent classification systems. The very same adaptation technique may be employed to improve the existing model for non-stationary new data.
There are at least two other ways of exploiting the existing labeled data from a similar application.
Adaptation may begin with an existing classification model. Using labeled data from a target application, a new model may be built based on the existing classification model. This method is similar to incorporating prior knowledge or exploiting unlabeled utterances for Boosting. In previous works, a model which fit both the training data and the task knowledge or machine labeled data was trained. In implementations consistent with the principles of the invention, a model that fits both a small amount of application specific labeled data and the existing model from a similar application may be trained. More formally, the Boosting algorithm tries to fit both the newly labeled data and the prior model using the following loss function:
is the Kullback-Leibler divergence (or binary relative entropy) between two probability distributions p and q. In implementations consistent with the principles of the invention, the probability distributions may correspond to the distribution from the prior model P(Yi[l]=1|xi) and to the distribution from the constructed model ρ(ƒ(xi,l)), where ρ(x) is the logistic function 1/(1+e−x). This term is basically the distance from the existing model to the new model built with newly labeled data. In the marginal case, if these two distributions are always the same, then the KL term will be zero and the loss function will be exactly the same as the first term, which is nothing but the logistic loss. Here, η is used to control the relative importance of these two terms. This weight may be determined empirically on a held-out set.
Note that most classifiers support a way of combining models or augmenting the existing model, so although this particular implementation is classifier (i.e. Boosting) dependent, the idea is more general. For example, in implementations that use a Naive Bayes classifier, adaptation may be implemented as linear model interpolation or Bayesian adaptation (like MAP) may be employed.
As an extension of this adaptation method, in some implementations consistent with the principles of the invention, adaptation may be combined with active learning. Active learning aims to minimize the number of labeled utterances by automatically selecting the utterances that are likely to be most informative for labeling. Thus, the existing model may be used to selectively sample the utterances to label for the target application, and do the adaptation using those utterances. This technique may eliminate the labeling of the examples or classes which are already covered by the existing model. It may be especially important to determine the initial set of examples to label when the labeling resources are scarce.
Since there is a previous model that may be used to obtain confidence scores for the examples from the target application, certainty-based active learning may be employed.
The adaptation method was evaluated using utterances from a database of a commercial system. Two applications were selected, T1, and T2, both from a telecommunications domain, where users have requests about their phone bills, calling plans, etc. T1 is a concierge-like application which has all the intents that T2 covers. T2 is used only for a specific subset of intents. The data properties are shown in Table 1. As seen the call-type perplexity (computed using the prior distributions of the intents) of T2 is significantly lower while the utterances are longer. T1 has about 9 times more data than T2. All the data is transcribed. Tests were performed using the Boostexter tool. For all experiments, word 12-grams were used as features. In order not to deal with finding the optimal iteration numbers, many iterations were performed, the error rate was obtained after each iteration and the best error rate was used in all of the results below.
Training Data Size
Test Data Size
Number of Intents
Average Utterance Length
In this experiment, the goal is to adapt the classification model for T1 using T2 so that the resulting model for T2 would perform better. Table 2 presents the baseline re-suits using training and test data combinations. The rows indicate the training sets and columns indicate the test sets. The values are the classification error rates, which are the ratios of the utterances for which the classifier's top scoring class is not one of the correct intents. The third row is simply a concatenation of both training sets (indicated by simple). The fourth row (indicated by tagged) is obtained by training the classifier with an extra feature indicating the source of that utterance, either T1 or T2. The performance of the adaptation is shown in the last 3 rows (indicated by adapt). As seen, although the two applications are very similar, when the training set does not match the test set, the performance drops drastically. Adding T1 training data to T2 does not help, actually it hurts significantly. This negative effect disappears when we denote the source of the training data, but no improvement has been observed on the performance of the classification model for T2. Adaptation experiments using different η values indicate interesting results. By using a value of 0.1, it is actually possible to outperform the model performance trained using only T2 training data.
Adaptation results for the experiments. “simple” indicates
simple concatenation, “tagged” indicates using an extra
feature denoting the source of training data, “adapt” indicates
adaptation with different η values.
adapt(η = 0.1)
adapt(η = 0.5)
adapt(η = 0.9)
Because the proposed adaptation method is expected to work better with less application specific training data, the learning curves are drawn as presented in
Embodiments within the scope of the present invention may also include computer-readable media for carrying or having computer-executable instructions or data structures stored thereon. Such computer-readable media can be any available media that can be accessed by a general purpose or special purpose computer. By way of example, and not limitation, such computer-readable media can comprise RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other medium which can be used to carry or store desired program code means in the form of computer-executable instructions or data structures. When information is transferred or provided over a network or another communications connection (either hardwired, wireless, or combination thereof) to a computer, the computer properly views the connection as a computer-readable medium. Thus, any such connection is properly termed a computer-readable medium. Combinations of the above should also be included within the scope of the computer-readable media.
Computer-executable instructions include, for example, instructions and data which cause a general purpose computer, special purpose computer, or special purpose processing device to perform a certain function or group of functions. Computer-executable instructions also include program modules that are executed by computers in stand-alone or network environments. Generally, program modules include routines, programs, objects, components, and data structures, etc. that perform particular tasks or implement particular abstract data types. Computer-executable instructions, associated data structures, and program modules represent examples of the program code means for executing steps of the methods disclosed herein. The particular sequence of such executable instructions or associated data structures represents examples of corresponding acts for implementing the functions described in such steps.
Those of skill in the art will appreciate that other embodiments of the invention may be practiced in network computing environments with many types of computer system configurations, including personal computers, hand-held devices, multi-processor systems, microprocessor-based or programmable consumer electronics, network PCs, minicomputers, mainframe computers, and the like. Embodiments may also be practiced in distributed computing environments where tasks are performed by local and remote processing devices that are linked (either by hardwired links, wireless links, or by a combination thereof) through a communications network. In a distributed computing environment, program modules may be located in both local and remote memory storage devices.
Although the above description may contain specific details, they should not be construed as limiting the claims in any way. Other configurations of the described embodiments of the invention are part of the scope of this invention. For example, hardwired logic may be used in implementations instead of processors, or one or more application specific integrated circuits (ASICs) may be used in implementations consistent with the principles of the invention. Further, implementations consistent with the principles of the invention may have more or fewer acts than as described, or may implement acts in a different order than as shown. Implementations consistent with the principles of the invention may include other classification tasks, such as topic classification or named entity extraction. Accordingly, the appended claims and their legal equivalents should only define the invention, rather than any specific examples given.
|Cited Patent||Filing date||Publication date||Applicant||Title|
|US5918222 *||Mar 15, 1996||Jun 29, 1999||Kabushiki Kaisha Toshiba||Information disclosing apparatus and multi-modal information input/output system|
|US6311159 *||Oct 5, 1999||Oct 30, 2001||Lernout & Hauspie Speech Products N.V.||Speech controlled computer user interface|
|US6532444 *||Oct 5, 1998||Mar 11, 2003||One Voice Technologies, Inc.||Network interactive user interface using speech recognition and natural language processing|
|US7139717 *||Oct 11, 2002||Nov 21, 2006||At&T Corp.||System for dialog management|
|US7167832 *||Oct 11, 2002||Jan 23, 2007||At&T Corp.||Method for dialog management|
|US20020135618 *||Feb 5, 2001||Sep 26, 2002||International Business Machines Corporation||System and method for multi-modal focus detection, referential ambiguity resolution and mood classification using multi-modal input|
|US20020152068||Feb 22, 2001||Oct 17, 2002||International Business Machines Corporation||New language context dependent data labeling|
|US20030036903 *||Aug 16, 2001||Feb 20, 2003||Sony Corporation||Retraining and updating speech models for speech recognition|
|US20030105634 *||Oct 11, 2002||Jun 5, 2003||Alicia Abella||Method for dialog management|
|US20040117759 *||Feb 21, 2002||Jun 17, 2004||Rippert Donald J||Distributed development environment for building internet applications by developers at remote locations|
|US20050091050 *||Mar 8, 2004||Apr 28, 2005||Surendran Arungunram C.||Systems and methods that detect a desired signal via a linear discriminative classifier that utilizes an estimated posterior signal-to-noise ratio (SNR)|
|US20050228641 *||Mar 31, 2004||Oct 13, 2005||Microsoft Corporation||Language model adaptation using semantic supervision|
|US20060195321 *||Feb 28, 2005||Aug 31, 2006||International Business Machines Corporation||Natural language system and method based on unisolated performance metric|
|US20060212413 *||Oct 3, 2005||Sep 21, 2006||Pal Rujan||Classification method and apparatus|
|US20070026406 *||Aug 13, 2004||Feb 1, 2007||Iconix Pharmaceuticals, Inc.||Apparatus and method for classifying multi-dimensional biological data|
|US20070198573 *||Mar 28, 2007||Aug 23, 2007||Jerome Samson||Data classification methods and apparatus for use with data fusion|
|EP1280136A1||Jul 16, 2002||Jan 29, 2003||AT&T Corp.||Spoken language understanding that incorporates prior knowledge into boosting|
|1||G. Di Fabbrizio, G. Tur, D. Hakkani-Tur, "Bootstrapping spoken dialog Systems with Data Reuse", Proceedings of SIGDIAL-2004, 5th SIGDIAL Workshop on Discourse and Dialogue, May 1, 2004, pages.|
|Citing Patent||Filing date||Publication date||Applicant||Title|
|US20140180692 *||Feb 19, 2014||Jun 26, 2014||Nuance Communications, Inc.||Intent mining via analysis of utterances|
|Mar 21, 2005||AS||Assignment|
Owner name: AT&T CORP., NEW YORK
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:TUR, GOKHAN;REEL/FRAME:016407/0880
Effective date: 20050310
|Dec 31, 2014||FPAY||Fee payment|
Year of fee payment: 4
|Mar 28, 2016||AS||Assignment|
Owner name: AT&T PROPERTIES, LLC, NEVADA
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:AT&T CORP.;REEL/FRAME:038275/0238
Effective date: 20160204
Owner name: AT&T INTELLECTUAL PROPERTY II, L.P., GEORGIA
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:AT&T PROPERTIES, LLC;REEL/FRAME:038275/0310
Effective date: 20160204